Uploaded by Bug Deer

The Hopcroft-Tarjan Planarity Algorithm

advertisement
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/228993896
The Hopcroft-Tarjan Planarity Algorithm
Article · November 1993
CITATIONS
READS
0
3,207
1 author:
William Kocay
University of Manitoba
98 PUBLICATIONS 659 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Reconstruction View project
Configurations View project
All content following this page was uploaded by William Kocay on 20 October 2014.
The user has requested enhancement of the downloaded file.
October, 1993
The Hopcroft-Tarjan Planarity Algorithm
William Kocay*
Computer Science Department
University of Manitoba
Winnipeg, Manitoba, CANADA, R3T 2N2
e-mail: bkocay@cs.umanitoba.ca
Abstract
This is an expository article on the Hopcroft-Tarjan planarity algorithm. A graph-theoretic analysis of a version of the algorithm is presented.
An explicit formula is given for determining which vertex to place first in the
adjacency list of any vertex. The intention is to make the Hopcroft-Tarjan
algorithm more accessible and intuitive than it currently is.
1. Introduction
Let G be a simple 2-connected graph with vertex set V (G) and edge set
E(G). The number of vertices of G is denoted by n. If u, v ∈ V (G), then
A(u) denotes the set of all vertices adjacent to u. v → u means that v is
adjacent to u, so that v ∈ A(u) (and since G is undirected, also u → v).
Refer to the book by Bondy and Murty [3] for other graph-theoretic terminology. Hopcroft and Tarjan [13] gave a linear-time algorithm to determine
whether G is planar, using a depth-first search. The depth-first search computes two low-point arrays for the vertices, L1(u) and L2 (u), and assigns a
weight to the edges of G, where the weight is computed from the low-points.
The edges incident on each u ∈ V (G) are then ordered according to their
weights. The algorithm then uses the revised graph in which incident edges
have been ordered, and performs another depth-first search, the so-called
PathFinder to embed the graph in the plane.
The purpose of this article is to explain exactly why the particular
weighting scheme used for the edges is necessary, to elucidate why it works,
and to suggest another, more intuitive way of embedding the graph in
the plane, once the low-point depth-first search has been executed. The
Hopcroft-Tarjan algorithm is complicated and subtle. For example, see
the description of it in the paper of Williamson [20]. In [2], Di Battista,
Eades, Tamassia, and Tollis state that “The known planarity algorithms
* This work was supported by an operating grant from the Natural Sciences and Engineering Research Council of Canada.
1
that achieve linear time complexity are all difficult to understand and implement. This is a serious limitation for their use in practical systems.
A simple and efficient algorithm for testing the planarity of a graph and
constructing planar representations would be a significant contribution”.
It may not be possible to construct a simple planarity algorithm, but the
graph-theoretic analysis of the algorithm presented here is intended to make
the algorithm easier to understand and implement. A number of textbooks
on graph theory and algorithms have been published since the paper of
Hopcroft and Tarjan, many of which discuss planarity algorithms. The
books by Even [9], Foulds [10], and Gould [12] all discuss the HopcroftTarjan algorithm, but none of them attempt to prove that the algorithm
works, because of the difficulty. The books by Gibbons [11], McHugh [16],
Bondy and Murty [3], and Chartrand and Oellermann [5] all choose to describe the planarity algorithm of Demoucron, Malgrange, and Pertuiset [7]
because it is conceptually much easier to understand than the HopcroftTarjan algorithm. The book by Nishizeki and Chiba [17] describes the
algorithm developed by Lempel, Even, Cederbaum [15], and Booth and
Lueker [4], because it also is more intuitive. In the book by Williamson
[19], the chapter on planarity ends with exercise 7.31. The reader is asked
to “try and convince at least two friends that your algorithm works in linear
time. A major hurdle will be the soporific effect on the friends”. It is very
easy to prove that a depth-first search based algorithm is linear. It is hard
to prove that the Hopcroft-Tarjan algorithm works. This paper intends to
make the Hopcroft-Tarjan algorithm more accessible and intuitive.
Consider a depth-first search of G. It constructs a rooted spanning
tree T of G, called a DF-tree of G. The root node is the vertex from which
the search was initiated. Each vertex v ∈ V (G), except the root, has a
unique ancestor vertex, being that vertex from which DFS(v) was called.
Denote it by a(v). Fig. 1 shows a DF-tree in a graph G. a(v) is the vertex
immediately above v in the tree. The depth-first search assigns a numbering
to the vertices, the depth-first numbering, written D(u), where u ∈ V (G).
It is the order in which the vertices are visited by the search. The vertices
in Fig. 1 are numbered according to their DF-numbering. The edges of the
tree are given by E(T ) = {uv | u = a(v)}.
DFS(u: vertex)
{ extend a DF-search in a graph G from vertex u }
begin
{ DFCount is a global counter initially set to 0 }
DFCount := DFCount + 1
D(u) := DFCount
2
for each v → u do begin
{ initially, all D(v) were set to 0 }
if D(v) = 0 then begin
a(v) := u
DFS(v)
end
end
end { DFS }
1
1
2
2
8
10
12
3
3
4
9
7
12
5
6
11
4
7
6
8
11
9
5
10
Fig. 1, A DFS in a graph
The tree T is drawn in Fig. 1 with its root at the top, the tree descends
from the root. The leaves of T are the vertices with no descendants. Starting at any vertex v in T , the sequence v0 = v, v1 = a(v0 ), v2 = a(v1), . . .,
is a sequence of vertices with decreasing DF-numbers, eventually reaching
the root vertex of T . The depth-first search divides the edges of G into two
categories. If uv is an edge with either a(v) = u or a(u) = v, then uv is a
tree edge. Otherwise uv is a frond .
Given any vertex u, write S(u) for the set of all descendants of u, that
is, all vertices v such that u occurs on the path from v to the root. Let
S ∗ (u) = {u} ∪ S(u). Let T (u) = {w ∈ A(v) | v ∈ S ∗ (u)}. In words, T (u)
consists of all vertices adjacent to u or any descendant of u. If a(v) = u,
the branch at u containing v is Bu(v) = {wx | w ∈ S ∗ (v)}, namely the set
of all edges incident on v or any descendant of v. The edge uv ∈ Bu (v) is
called the stem of the branch. At this point it is convenient to follow [13] or
[9] and renumber the vertices so that D(u) = u, that is, to refer to vertices
directly by their DF-number. We then define
L1(u) = min T (u)
L2(u) = min T (u) − {L1 (u)}
Since G is 2-connected, every vertex has degree at least 2, so that
|T (u)| ≥ 2. It follows that the low-points are well-defined. The algorithm
3
cannot refer to vertices directly by their DF-number, but must work with
the graph as it is given. So the algorithm must compute the low-points as
L1 (u) = x ∈ T (u), such that D(x) = min{D(v) | v ∈ T (u)}
L2 (u) = x ∈ T (u), such that D(x) = min{D(v) | v ∈ T (u) − {L1 (u)}}
However, it is much easier to decsribe the algorithm if vertices can be referred to via their DF-number. This means that all comparisons of vertices
such as “if v < w . . .” must be evaluated via the DF-number, namely “if
D(v) < D(w) . . .”, etc.
It is easy to modify the DFS to compute L1 (u), and only slightly more
difficult to modify it to compute L2 (u) as well. See [13] or [9] for details.
Suppose that u = a(v). Then u ∈ T (v). Therefore L1 (v) ≤ u. It is a classic
result of Hopcroft and Tarjan [1] that u is a cut-vertex of G if L1 (v) = u,
when u is not the root node. Since G is assumed to be a 2-connected graph,
it has no cut-vertices, so we can conclude that L1(v) < u whenever u = a(v)
and u is not the root node. It follows that L2(v) ≤ u.
2. Ordering the Edges
The graph G is stored by adjacency lists, namely Adj[u] is a linked list
containing all vertices adjacent to u. So Adj[u] is an ordered list of adjacent
vertices. If v ∈ A(u), then the node in Adj[u] corresponding to v represents
the edge uv ∈ E(G).
Hopcroft and Tarjan assign a weight to each edge. The weight is stored
as a field in each node of the linked list. wtu [v] denotes the weight of edge
uv in Adj[u]. It can be defined as follows. Suppose that v ∈ A(u). wtu [v]
is computed by DFS(u), by adding several statements.

2v,


2L1 (v),
wtu [v] =
 2L1 (v) + 1,

2n + 1
if uv is a frond with v < u
if a(v) = u and L2 (v) = u
if a(v) = u and L2 (v) < u
otherwise
The low-point depth-first search computes D(u), a(u), L1 (u), L2 (u),
and wtu [v], for all u ∈ V (G) and all v ∈ A(u). Call this first DFS the
LowPtDFS . Following it the edge-weights are in the range 1..2n + 1. The
edges are then sorted by a bucket sort, and the adjacency lists Adj[u] are
reconstructed in order of increasing wtu [v]. In the Hopcroft-Tarjan algorithm a second depth-first search is then called, the PathFinder . It uses
the re-ordered adjacency lists to recursively generate paths in the graph
and adds the paths one by one to build a planar embedding of G. The
algorithm is based on paths.
A different method is here proposed of embedding G, based on embedding the DF-tree T branch by branch rather than by embedding paths. The
4
reason for this is that the structure of trees is recursive in terms of their
branches. This recursion can be used to give a graph-theoretic proof by induction that the algorithm works, using the branches of T . The algorithm
is very similar to the path-embedding algorithm, but the use of branches
seems simpler and is mathematically advantageous. The first thing is to
understand the purpose and function of the weights.
r
a
b
Bu (v)
v
Bu (w)
u
w
s
x
t
Bu (x)
c
Fig. 2, Branches and fronds at u in a DF-tree T
Consider a DF-tree T constructed by the LowPtDFS , drawn schematically in the plane, as in Fig. 2, where only some of the vertices and edges
are shown. The tree edges are shown in bold. At u there are two fronds,
ua and ub, and there are three branches Bu (v), Bu (w), Bu (x) with stems
uv, uw, and ux. The fronds ua and ub have weights wtu [a] = 2a and
wtu [b] = 2b, since frond-weights are determined by their endpoints. To
compute the weight of the stems, notice that L1 (x) = r and L2 (x) = u, so
that wtu [x] = 2r, where r is the root vertex. L1 (v) = a and L2 (v) = b < u,
so wtu [v] = 2a+1. L1 (w) = b and L2 (w) = u, so wtu [w] = 2b. The graph is
to be embedded by another DFS which traverses T recursively, embedding
the branches and fronds. The embedding takes place from the outside in. At
vertex u they are embedded in the order Bu(x), ua, Bu (v), ub, and Bu (w),
since wtu [x] = 2r < wtu [a] = 2a < wtu [v] = 2a+1 < wt u[b] = 2b = wtu [w].
Embedding a branch like Bu (v) with L1(v) = a is very much like
embedding a frond ua. In fact if we insert a vertex of degree two on the
frond ua, it behaves exactly like a branch whose L1 -value equals a. This
is why the weight of a frond is determined by its endpoint, and that of a
branch by the low-point of its stem vertex. Because L2 (v) < u there will
5
be other fronds inside the branch between u and a. Thus we can not know
while visiting u whether it would be possible to embed the frond ua after
Bu [v] has been embedded. Therefore the frond must be embedded before
the branch. This is why the frond is given weight 2a and the stem weight
2a + 1.
A branch like Bu (w) has L2 (w) = u. Therefore after the branch has
been embedded it will still be possible to embed the frond ub, where b =
L1 (w) inside the branch. The order in which the frond and branch should
be embedded is still not determined. Therefore they are given equal weight
2b. So branches are ordered according to the L1 -value of the stem vertex.
The purpose of the L2 -value is to distinguish the two types of branches. The
factor of 2 in the weight computation is to distinguish between branches
whose L2 -value is less than u, or equal to u, by making their weight odd or
even, respectively.
2.1 Definition. A branch Bu (v), where u = a(v), with L2 (v) = u is
called a type I branch. If L2 (v) < u it is called a type II branch.
So at vertex u, the ordering of Adj[u] defined by the weights places
fronds and type I branches before type II branches with the same L1value. A frond behaves like a type I branch. The definition of wt assigns
wtu [v] = 2n + 1 for some edges uv. These are either downward fronds,
that is, fronds where v > u, or else tree edges with a(u) = v. These are
embedded from their other endpoint, and so can be ignored while visiting
u. Thus they are given a weight which places them after all other fronds
and branches.
3. The Branch Points of T
Let T be a DF-tree constructed by a DFS. The DFS starts at the root of
T . It descends the tree by recursively calling itself. While visiting node u,
DFS(v) may be called recursively for several v ∈ Adj[u].
3.1 Definition. A vertex u is a branch point of T , if T has a branch
Bu (v) where v ∈ Adj[u] is not the first vertex of Adj[u]. The root of T is
also a branch point. Given any v in T we define b(v), the branch point of
v as follows.
1) b(v) = v, if v is the root of T ;
2) if v is not the root, then let u = a(v).
n
b(v) = b(u), if v is the first node of Adj[u];
u
otherwise.
So the branch points of a given DF-tree T depend on which node comes
first in the adjacency lists. Since the sequence u1 = a(v), u2 = a(u1), . . .
eventually reaches the root, b(v) is uniquely defined, and is always a branch
point of T . It is easily computed by a DFS initiated from the root of T .
6
DFS(u: vertex)
{ DFS to compute the branch points of an existing DF-tree T }
{ b(u) is already known }
begin
FirstTime := true { indicates first v ∈ Adj[u] }
{ a(u), D(u) have been computed by a previous DFS }
for each v ∈ Adj[u] do begin
if a(v) = u then begin
{ descend the tree }
if FirstTime then b(v) := b(u)
else b(v) := u
DFS(v)
end
FirstTime := false
end
end { DFS }
The LowPtDFS cannot compute the branch points. This is because the
branch points depend on the ordering of Adj[u], and this is not determined
until after the LowPtDFS has been executed. Given a DF-tree T and
any ordering of the adjacency lists, a set of branch points of T is then
determined. A subsequent DFS using the reordered Adj[u] will descend the
same tree T , although the branches will be searched in an order determined
by Adj[u].
Let L(T ) = {v | the first w ∈ Adj[v] corresponds to a frond vw}.
Clearly L(T ) contains all the leaves of T . It may also contain other vertices
which are not leaves. If v ∈ L(T ) is not a leaf, then v is always a branch
point of T . For each v ∈ L(T ), let Pv denote the path in T from v to b(v).
The length of the path is `(Pv ), the number of edges in the path.
Referring to Fig. 2, the branches at u were ordered by the weights as
Bu (x), Bu(v), and Bu (w). The branch points are then as follows: b(x) = r,
b(v) = u, b(w) = u, b(u) = r, b(c) = r, etc. The branch points divide the
tree into paths. In Fig. 2, L(T ) consists of the leaves {c, s, t}, and the paths
Pc , Ps , and Pt together contain all of T .
3.2 Lemma. Let T be a DF-tree of G, with branch points defined by the
ordered adjacency lists of G. Then {Pv | v ∈ L(T )} is a partition of T into
paths.
Proof . By induction on |L(T )|. Let r denote the root of T . If L(T ) = {v},
then v is a leaf and T = Pv , a path from v to r. Suppose that |L(T )| =
k, where k > 1. Let v be the last leaf visited, and let u = b(v). Pick
w ∈ Pv such that a(w) = u. Then T 0 = T − Pv is a DF-tree for the
graph G − S ∗ (w) with |L(T 0 )| = k − 1. Clearly L(T 0 ) = L(T ) − v. Hence
7
{Px | x ∈ L(T 0 )} ∪ {Pv } is a partition of T into paths. The lemma follows.
At vertex u in T , the first node v ∈ Adj[u] is crucial to the success of
the algorithm. This is why the branch points are important. When uv is
a tree edge, the DFS continues to descend the tree. When uv is a frond
the descent stops. In this case the path Pu in T from u to b(u), plus the
frond uv is the path found by the PathFinder DFS of Hopcroft and Tarjan
[13]. They also consider each frond as a path of length one. Hopcroft and
Tarjan embed G in the plane by embedding these paths one by one.
We view the algorithm from a slightly different perspective. We use
a DFS to embed G branch by branch of T . A frond uv, where v < u, is
embedded either on the left side of T or on the right side. So we have an
embedding of T in the plane, with the fronds arranged around T giving
an embedding of G. We first determine an ordering of Adj[u] so that the
branch points are guaranteed to permit an embedding when G is planar.
Then, following Hopcroft and Tarjan, we keep two linked lists of fronds, LF
and RF containing the fronds and branches embedded on the left of T and
on the right of T , respectively. LF and RF can be viewed as stacks, whose
tops contain the number of the vertex currently marking the upper limit
in the tree to which fronds may be embedded. As a first approximation we
have
EmbedBranch(u: vertex)
begin
{ NonPlanar is a global flag }
for each v ∈ Adj[u] do begin
NonPlanar := true { assume G will be found non-planar }
if a(v) = u then begin { uv is a tree edge }
if b(v) = u then begin
{ uv begins a new branch at u }
if L1(v) is too small to permit an embedding then Exit
place Bu (v) either on LF or RF
end
EmbedBranch(v)
if NonPlanar then Exit
end
else if v < u then begin { uv is a frond }
EmbedFrond(u, v)
if EmbedFrond is unsuccessful then Exit
end
end
NonPlanar := false { no conflicting fronds were found }
end { EmbedBranch }
8
EmbedFrond is a procedure which attempts to embed the frond uv by
placing it either on LF or RF . It returns a boolean value indicating whether
uv was successfully embedded.
4. Ordering the Adjacency Lists
The ordering of Adj[u] by weights is not sufficient to guarantee that an
embedding is possible without further refinement. We present several examples.
1
1
2
2
3
7
7
6
5
4
3
5
4
6
Fig. 3, Example 1
Example 1 shows two attempts at embedding the same graph. The tree
T is the same in both attempts, but the ordering of the branches differs. In
the example on the left, the adjacency lists are ordered strictly according
to the weights computed. On the right the ordering has been adjusted
slightly. The branches and fronds are the same in each case. At vertex
3, B3 (4) is a type I branch with wt3 [4] = 2. There is a frond (3,1) with
wt3 [1] = 2. In the left example, the frond was embedded first, then the
branch, thereby forcing the branch inside the frond on the left side of T .
In the right example, the branch B3 (4) was embedded before the frond. At
vertex 4, there is a type II branch B4 (5) of weight 3. There is also a frond
(4,1) of weight 2. On the left, the frond is embedded first; on the right the
branch is taken first. At vertex 5, there are two type II branches, B5 (6)
and B5 (7). Both have weight 3. There is also a frond (5,1) of weight 2.
On the left, the frond is taken first, then B5 (6), then B5 (7); on the right
B5 (6) is taken first, then the frond, then B5 (7). The thing to notice is that
in the example on the left, it is impossible to embed B5 (7) without forcing
a crossing of the edge (6,3). So the ordering of Adj[u] by weights must be
adjusted if this graph is to be successfully embedded.
9
1
1
2
3
2
3
7
5
6
5
4
4
6
7
Fig. 4, Example 2
Example 2 also shows two attempts at embedding a graph. At vertex
3 there are two type II branches, B3 (4) and B3(5), of weight 3. In both
cases B3 (4) is taken first. At vertex 5 there are also two type II branches,
B5 (6) and B5 (7), both of weight 3. On the left B5 (6) is taken first. We
then find that B5 (7) cannot be embedded without crossing the edge (6,2).
On the right B5(7) is taken first. Notice that in the diagram on the left,
moving B5 (7) to the other side of the path from 3 to 6 is not permissible,
since that would change the branch points of the tree. This is equivalent
to reordering the adjacency lists, which could then affect the portion of the
branch B3 (5) already traversed at that point.
Example 3 shows a graph successfully embedded in the plane. At vertex 3 there is a type I branch B3 (4) with wt3 [4] = 2 and a type II branch
B3 (5) with wt3 [5] = 3. The type I branch is taken first and the type II
branch is then embedded inside the face created by the frond (4,1). At vertex 9 there is a type I branch B9 (10) and a type II branch B9 (11). Again
the type I branch is taken first. At vertex 15 there is a frond (15,1) with
wt15 [1] = 2 and a type II branch B15(16). The frond is taken first. Returning up the tree a number of minor branches are embedded, all successfully.
The thing to notice is that at vertex 15, if the branch B15(16) had been embedded before the frond, then it would have been impossible to embed the
minor branches when returning up the tree without introducing a crossing.
So the frond must come first.
10
1
2
5
19
16
15 14
3
11 9
18
17
4
10
Fig. 5, Example 3
Example 4 shows a very similar graph successfully embedded in the
plane. As in example 3 there are two branches at vertex 3, of which the
type I branch is taken first. At vertex 9 there are again two branches and
the type I branch is taken first. At vertex 15 there is a frond (15,1) and
a type II branch B15 (16). This time the branch is taken first, and the
minor branches are then succesfully embedded when returning up the tree.
The branch must be embedded before the frond. If the frond (15,1) were
taken first it would not be possible to embed the minor branches without
introducing a crossing.
1
2
16 15
5
19
17
3
11 9
14
18
10
Fig. 6, Example 4
11
4
4.1 Definition. Let T be a DF-tree in a graph G. We define two integer
functions N (v) and N 0 (v) for every v ∈ V (G) as follows.
1) If v is the root of T then N (v) = N 0 (v) = 0.
Otherwise let u = a(v).
2) If Bu(v) is a type I branch then N (v) = N 0 (v) = 0.
3) If Bu(v) is a type II branch, choose the unique vertex x such that
a(x) = L2(v). Then L2 (v) < x ≤ u < v.
3.1) If b(v) = 1 then N (v) = 0.
If L2 (v) < b(v) then N (v) = 1.
Otherwise L2 (v) ≥ b(v) 6= 1. Define

 N (x) + 1, if N (x) 6= 0
N (v) = 2,
if N (x) = 0 and u → L1 (v)

N (u),
if N (x) = 0 and u 6→ L1 (v)
3.2) If b(u) = 1 then N 0 (v) = 0.
If L2 (v) < b(u) then N 0 (v) = 1.
Otherwise L2 (v) ≥ b(u) 6= 1. Define

 N (x) + 1, if N (x) 6= 0
N 0 (v) = 2,
if N (x) = 0 and u → L1 (v)

N (u),
if N (x) = 0 and u 6→ L1 (v)
The definition of N (v) and N 0 (v) appears complicated, but is only a
mathematical formulation of a simple idea which is necessary in order to
guarantee that the algorithm works. It can be computed quite easily, just by
following the definition. N (v) counts the number of steps needed to reach
a vertex smaller than b(v), where the steps are based on fronds to L2 (v) or
L1 (v), as described below. If this is not possible then N (v) = 0. If b(v) = 1
then there is no vertex smaller than b(v), so N (v) = 0. N 0 (v) counts the
number of steps needed to reach a vertex smaller than b(u). If v is the first
node in Adj[u], then b(v) = b(u), so N (v) = N 0 (v). Consider the example of
Fig. 6. N (1) = 0. L2 (2) = 1, so N (2) = 0. L2 (3) = 2 ≥ b(3), so N (3) = 0.
Similarly N (4) = 0. L2 (5) = 2 < b(5) = 3, so N (5) = 1. L2(6) = 2 <
b(6) = 3, so N (6) = 1. L2 (7) = 5 ≥ b(7) = 3, so N (5) = 1 + N (6) = 2, since
a(6) = 5. Continuing in this way, we have N (8) = 2, N (9) = 3, N (10) = 0,
N (11) = N (12) = N (13) = N (14) = 1, N (15) = N (16) = 2. We see that
N (v) counts the number of steps needed to reach a vertex smaller than b(v).
If L2 (v) < b(v), then one step is sufficient, that is, the L2 -frond in Bu (v)
joins to a vertex < b(v); hence N (v) = 1. If Bu (v) is a type I branch then
it is not possible to reach a vertex smaller than b(v) via the L2 function;
hence N (v) = 0 in this case. Otherwise Bu (v) is a type II branch with
L2 (v) ≥ b(v); for example in Fig. 6, L2 (7) = 5 > b(7) = 3. In this case,
12
pick x where a(x) = L2 (v); in this example we get x = 6. Continuing
with the example, starting at vertex 7, we can use two L2 -fronds, (18,5)
and (19,2), giving a path (7,8,18,5,6,19,2) to reach vertex 2 < b(7) = 3.
Starting at vertex 9, we can use three L2 -fronds, (12,7), (18,5), and (19,2)
to reach vertex 2 < b(9) = 3. So N (7) = 2 and N (9) = 3.
N 0 (v) differs from N (v) only in the use of b(u) in the comparison
instead of b(v). So N (v) = N 0 (v) when v is the first node in the adjacency
list. When v is not the first node, N 0 (v) is the value that N (v) would have,
if v were the first node. N 0 (v) is used to determine the correct ordering of
Adj[u]. It is really N 0 (v) that we want, but since N 0 (v) is defined in terms
of N (v), we must compute N (v) as well.
The functions N (v) and N 0 (v) can easily be computed by adding several statements modelled on definition 4.1 to the DFS of section 3 that
computes the branch points. N (v) and N 0 (v) depend on the branch points
of those vertices of T which are ancestors of v. The branch points in turn
depend on the ordering of the adjacency lists. N 0 (v) is used to re-order the
adjacency lists. Initially, Adj[u] is ordered according to the weight function
wtu [v]. Furthermore we assume that if v, w ∈ Adj[u] have equal weight,
where uv is a frond, and uw is the stem of a branch, that v precedes w
in the adjacency list. That is, we assume that fronds precede branches of
equal weight. This can easily be accomplished in the bucket sort by placing
fronds at the head of a bucket and branch-stems at the tail of a bucket.
Let v∗ be the first vertex of Adj[u], and let W be the smallest odd
number ≥ wtu [v ∗ ]. Let A∗ (u) = {v ∈ Adj[u] | wtu [v] ≤ W }. A∗ (u)
represents all type I and type II branches at u with L1-value = bW/2c (and
possibly a frond at u to vertex bW/2c). We now compute N 0 (v) for each
v ∈ A∗ (u) and adjust the ordering of Adj[u] according to the following rule.
4.2 Rule.
1) If there is a v ∈ A∗ (u) such that Bu (v) is a type II branch for which
N 0 (v) is even, then v is moved to the head of Adj[u]. (If there is more
than one such v, only one of them is moved.)
2) Otherwise, if there is a v ∈ A∗ (u) such that Bu (v) is a type I branch,
then v is moved to the head of Adj[u]. (If there is more than one such
v, only one of them is moved.)
3) Otherwise Adj[u] is unchanged.
The computation of N 0 (v) and the adjustment of the ordering of Adj[u]
is accomplished by the DFS which computes the branch points of T , since
the ordering of the adjacency lists affects the branch points. If x is the
unique vertex such that a(x) = w, for some vertex w, then we need to be
able to find x, given w. Edge wx is the stem of the branch Bw (x) at w.
The easiest way to find x is for the algorithm to maintain an array Stem[w],
13
for vertices w.
BranchPtDFS(u: vertex)
{ DFS to compute b(v) and N (v), for all v ∈ Adj[u], and to }
{ reorder Adj[u]. Upon entry b(w) and N (w) are known, }
{ where w is u or any ancestor of u }
begin
v := first vertex of Adj[u]
W := wtu [v]
if W is even then W := W + 1
vI := 0; vII := 0 { markers for type I and II branches }
for each v ∈ Adj[u] such that wt u[v] ≤ W do begin
if a(v) = u then begin { uv is a tree edge }
N (v) := 0 { assume N (v) will be 0 }
if wtu [v] is even then vI := v { Bu (v) is a type I branch }
else begin { Bu (v) is a type II branch }
if L2(v) < b(u) then N (v) := 1
else if b(u) 6= 1 then begin
x := Stem[L2 (v)]
if N (x) 6= 0 then N (v) = N (x) + 1
else if u → L1 (v) then N (v) := 2
else N (v) = N (u)
end
if N (v) is even then begin vII := v; go to 1; end
end
end { if a(v) = u }
end { for }
1: if vII 6= 0 then put vII at head of Adj[u]
else if vI 6= 0 then put vI at head of Adj[u]
{ Adj[u] is now reordered, descend T }
FirstTime := true { indicates first v ∈ Adj[u] }
for each v ∈ Adj[u] do begin
if a(v) = u then begin
b(v) := u { assume the branch point will be u }
if FirstTime then b(v) := b(u)
else if wtu [v] is even then N (v) := 0 { type I branch }
else N (v) := 1 { type II branch }
Stem[u] := v { assign stem for branch about to be searched }
BranchPtDFS(v)
end
FirstTime := false
end
end { BranchPtDFS }
14
4.3 Lemma. The BranchPtDFS computes b(v) and N (v) correctly, for
all v ∈ V (G).
Proof . By induction on the number of levels of recursion. Since we have
identified vertices with their DF-number, the root is vertex 1. We must
assign b(1) := 1 and N (1) := 0 before calling BranchPtDFS(1). So it is
true with 0 levels of recursion. Suppose it is true up to k levels of recursion,
and let u be a vertex visited at the (k − 1)st level of recursion, where k ≥ 1.
The first for-loop of the algorithm takes each v ∈ A∗ (u) in turn, if uv is a
tree edge. It initiallizes N (v) := 0. If Bu (v) is a type I branch, it leaves
N (v) with the value of 0, which is correct. It records the vertex v in the
value vI . If Bu (v) is a type II branch, it compares L2 (v) with b(u). If L2 (v)
is smaller, it sets N (v) = 1. This is the value of N 0 (v). If b(u) = 1, the
value of N (v) is left as 0. Otherwise, if L2 (v) is larger than b(u), it finds x
as in definition 4.1. The value given to N (v) is then one of N (x) + 1, 2, or
N (u), as in 4.1. Again this is the value of N 0 (v). If N 0 (v) is even, it records
the vertex v in the value vII . So after the for-loop has been executed, N (v)
contains the value of N 0 (v).
At this point one of vI or vII may be moved to the head of Adj[u]. Since
N (v) = N 0 (v) when v is the first node in Adj[u], N (v) has the correct value
for the first vertex in the list. It also sets b(v) = b(u) for the first node in
the list, if uv is a tree edge. All remaining tree edges uv have b(v) = u,
and N (v) = 0 or 1, according as they belong to type I or II branches,
respectively. These are the values assigned by the algorithm. It follows
that b(v) and N (v) are correctly computed for each v ∈ Adj[u] where uv is
a tree edge. It then calls BranchPtDFS recursively. Therefore the lemma
is true for every node visited at the k th level of recursion. By induction it
is true for all levels.
5. Graph-Theoretic Analysis of the Algorithm, |L(T )| = 1.
The re-ordering of Adj[u] according to rule 4.2 is exactly what is needed
in order to guarantee that the algorithm EmbedBranch will be successful
whenever G is planar. The proof is by induction on n (the number of
vertices of G), and on |L(T )|. We assume that G is a 2-connected graph
and that n ≥ 3.
Suppose first that |L(T )| = 1. Let L(T ) = {v}, and let r denote the
root of T . Since G is 2-connected, the first vertex in Adj[v] is r, representing
a frond to the root of T . Without loss of generality vr will be embedded
on the left side of T . The cycle C formed by vr and Pv divides the plane
into two regions, the inside and outside of C. The algorithm maintains two
linked lists of fronds, LF and RF (note that a linked list is a sequence).
Without loss of generality, we can assume that the left side of T refers to
the inside region and the right side to the outside region of C. Every frond
15
is to be placed either in LF or RF . Hopcroft and Tarjan use these linked
lists as stacks, and discard fronds that can no longer have an effect on the
remaining part of the algorithm. But it is advantageous to keep the entire
linked list. Let LF = (f1 , f2 , f3 , . . .) and RF = (g1 , g2 , g3, . . .). The fronds
are ordered according to their smaller endpoint, namely, if fi = ai bi where
ai < bi , then a1 ≤ a2 ≤ a3 ≤ . . ., and similarly for RF . As we shall see, the
sequences LF and RF constructed by the algorithm completely determine
the combinatorial embedding of G in the plane. Therefore when we say
that the algorithm embeds G in the plane, we mean that is constructs the
lists LF and RF .
5.1 Definition. If xy and wz are two fronds such that x < w < y < z
then they cannot both be on the same side of C. They are called conflicting
fronds. The frond graph of G and T is denoted F G. Its vertices are the
fronds of G. Two fronds are adjacent if they conflict.
5.2 Lemma.
If G is planar and |L(T )| = 1, then F G is a bipartite graph.
Proof . V (F G) can be partitioned into those fronds inside C and those
outside C. Since G is planar, there are no conflicting fronds both inside C
or both outside C. Hence F G is bipartite.
It follows that F G contains no odd cycles. In general the frond graph
will consist of a number of connected components F1 , F2 , . . ., Fm . (Hopcroft
and Tarjan call each Fi a block of fronds.) For each i, let (Li , Ri ) be the
bipartition of Fi , where Li consists of those fronds of Fi embedded on the
left side of T , and Ri those fronds embedded on the right side of T . The
bipartition (∪Li , ∪Ri ) of F G is not unique if m > 1, for we may exchange
Li and Ri in any Fi to obtain another bipartition of F G corresponding to
another embedding of G. This is what Hopcroft and Tarjan call switching
sides.
The diagram shows a DF-tree T for which F G has 4 components,
indicated by the shading of the lines. A nesting of the components is
evident.
16
Fig. 7, Components of the frond graph, |L(T )| = 1.
Initially F G = Ø. In order to avoid the situation when F G = LF =
RF = Ø it is convenient to create two “dummy” fronds f0 = (0, n) and g0 =
(0, n), and a “dummy” component F0 for which L0 = {f0 } and R0 = {g0 }.
The algorithm descends T to the leaf v. All fronds are embedded while
EmbedBranch is re-ascending T . As fronds are assigned either to LF or
RF , F G is gradually constructed. The first frond embedded is vr. Without
loss of generality, the algorithm will first try to place a frond on the left,
if possible. So the algorithm will set f1 = vr. Let ui denote the minimum
vertex of any frond in Li , namely ui = min{a | ∃ab ∈ Li }. Similarly vi =
max{a | ∃ab ∈ Li }, xi = min{a | ∃ab ∈ Ri }, and yi = max{a | ∃ab ∈ Ri }.
If Fi consists of a single frond, then Ri will be empty, so that xi and yi are
not always defined. In this case we take xi = n and yi = 0. In this way
most of the inequalities will be satisfied even when Ri = Ø. Every Fi will
have Li 6= Ø. In terms of the DF-numbering the first frond is f1 = (1, n).
Let u be a vertex of T being visited by EmbedBranch when a frond
uw is encountered, where w < u. Any component Fi such that ui , xi ≥ u
cannot possibly affect the remaining execution of the algorithm, since no
remaining frond can conflict with any frond of Fi . Therefore the algorithm
does not need to know all components of F G, only those for which u > ui
or u > xi . Let F0 , F1 , . . ., Fm denote all such components, in the order in
which they were constructed, where m ≥ 1. We can assume that u ≤ vi for
every i, and u ≤ yi for every i for which Ri 6= Ø, since fronds are embedded
while ascending T .
5.3 Lemma. Let 1 ≤ j ≤ m. Each component Fj “fits inside” Fj−1 ,
namely Fj−1 contains a frond ab such that a ≤ uj < vj ≤ b, and if Rj 6= Ø
17
then a ≤ xj < yj ≤ b.
Proof . The frond graph is a dynamic structure, changing as fronds are
embedded. By assumption F1 , . . ., Fm consists of those components for
which ui < u ≤ vi or xi < u ≤ yi , where u is the vertex currently being
visited by EmbedBranch. Wlog, let uj < u.
Suppose first that uj−1 ≤ uj . Then uj < u ≤ vj−1 . Fj−1 is connected,
so it contains a sequence of fronds connecting uj−1 to vj−1 and alternating
between Lj−1 and Rj−1 . Therefore a < uj < b for some frond ab of this
sequence in Fj−1 . It follows that vj ≤ b and that a ≤ xj < yj ≤ b if Rj 6= Ø
since Fj does not conflict with Fj−1 .
On the other hand, suppose that uj < uj−1 . Then since Fj does not
conflict with Fj−1 , uj < u ≤ vj ≤ uj−1 < vj−1. However at least one of
ui < u ≤ vi or xi < u ≤ yi holds for every i, so that xj−1 < u ≤ yj−1 must
hold. We can then apply the previous argument using xj−1 and yj−1 in
place of uj−1 and vj−1 .
Let ` denote the largest integer such that the smaller endpoint of f`
is < u and let `w denote the smaller endpoint. Similarly, let r denote the
largest integer such that the smaller endpoint of gr is < u and let rw denote
this endpoint. Because of F0 we can be sure that ` and r are always defined.
There are three possible situations that can arise for vertex u, as illustrated
below.
xm
rw
rw
um
rw
lw
u
um u
vm
ym
vm
um
xm
lw
u
xm
ym
vm
ym
Fig. 8, Situation of vertex u wrt Fm , cases 1, 2, and 3
1. u > um and u > xm .
In this case Fm contains a frond to um and xm , so u > `w ≥ um and
u > rw ≥ xm . `w and rw both belong to fronds of Fm .
2. u ≤ um and u > xm .
In this case um ≥ u > `w and u > rw ≥ xm . `w belongs to a frond of
Fi where i < m. rw belongs to a frond of Fm . By Lemma 5.3, `w ≤ rw .
18
3. u > um and u ≤ xm .
In this case u > `w ≥ um and xm ≥ u > rw . `w belongs to a frond
of Fm . rw belongs to a frond of Fi , where i < m. By Lemma 5.3,
rw ≤ ` w .
There are several possible situations that can arise for vertex w.
4. w ≥ `w and w ≥ rw .
This can occur with cases 1, 2, or 3 above. uw does not conflict with
any frond of LF or RF . uw is embedded on the left. It is inserted in
LR immediately after f` . A new component Fm+1 of F G is created
with um+1 = w and vm+1 = u. Rm+1 = Ø. The values of m and ` are
increased by 1.
5. w ≥ `w and w < rw .
This can only occur with cases 1 and 2 above, since `w < rw . uw
conflicts with a frond of Rm , but not Lm . uw is embedded on the
left. It is added to Lm , and inserted in LF immediately after f` . ` is
increased by 1. If case 2 applies then uw may conflict with a frond of
Fm−1 . Set um equal to w and execute the following steps:
while w < xm−1 do Merge(Fm−1, Fm )
Merge(Fm−1 , Fm ) is a procedure which updates the values um , vm , xm
and ym stored in the data structures, as follows.
if um < um−1 then um−1 := um
if vm > vm−1 then vm−1 := vm
if xm < xm−1 then xm−1 := xm
if ym > ym−1 then ym−1 := ym
discard Fm
m := m − 1
It also conceptually merges Lm−1 with Lm , and Rm−1 with Rm , although it is not actually necessary to store these sets explicitly.
6. w < `w and w ≥ rw .
This can only occur with cases 1 and 3 above, since rw < `w . uw
conflicts with a frond of Lm , but not Rm . uw is embedded on the
right. It is added to Rm , and inserted in RF immediately after gr . r
is increased by 1. If case 3 applies then uw may conflict with a frond
of Fm−1 . Set xm equal to w and execute the following steps:
while w < um−1 do Merge(Fm−1, Fm )
7. w < `w and w < rw .
This can occur with cases 1, 2, or 3 above. uw conflicts with a frond of
LF and RF . If case 1 applies then G is non-planar, since uw conflicts
with fronds f` and gr of Fm , thereby creating an odd cycle in Fm .
Execute the following steps:
19
while w < `w and w < rw do begin
if u > um and u > xm (case 1) then Exit (G is non-planar)
SwitchSides(Lm , Rm )
end
SwitchSides is a procedure that exchanges Lm and Rm and updates
the data structures as follows.
SwitchSides(Lm , Rm )
begin
if u ≤ um then begin { case 2 }
while um−1 > `w do Merge(Fm−1 , Fm )
`w := rw ; ` := r
adjust r so that gr is first frond preceding xm in RF
set rw to smaller endpoint of gr
end
else begin { u > um , case 3 }
while xm−1 > rw do Merge(Fm−1 , Fm )
rw := `w ; r := `
adjust ` so that f` is first frond preceding um in LF
set `w to smaller endpoint of f`
end
exchange the portion of the linked list LF between um and vm
with the portion of RF between xm and ym .
exchange um and xm , vm and ym , Lm and Rm
Merge(Fm−1 , Fm )
end { SwitchSides }
The while-loop will then stop with either w ≥ `w (case 5) or w ≥ rw
(case 6). In either case uw can then be embedded.
F1 , F2, . . ., Fm are connected components of the frond graph F G. If
uw conflicts with both Li and Ri , for some i, then uw cannot be embedded
without creating an odd cycle in Fi . Hence G is non-planar in this case.
Otherwise uw conflicts with at most one of Li and Ri , for each i. It is
then always possible to embed uw, for example by exchanging Li and Ri
for each i for which uw conflicts with Ri .
5.4 Lemma. Let |L(T )| = 1. If the algorithm runs to completion then G
is planar.
Proof . The embedding is built by adding one frond at a time to an existing
embedding, which is initially a planar embedding of T . Each frond is embedded without crossings, inside a face of the existing embedding. When
20
SwitchSides is executed, Lm and Rm are exchanged. This does not introduce any crossings into the embedding. So the embedding is always planar,
right up to completion.
It follows that if G is non-planar, the algorithm will not run to completion. In proving that the algorithm works, we can therefore assume that
G is planar.
5.5 Theorem. If G is planar and |L(T )| = 1, the algorithm will embed
G in the plane.
Proof . Suppose that the algorithm does not run to completion. So it stops
while visiting a vertex u with a frond uw, w < u, where w < `w and w < rw
(case 7). If case 1 applies (u > um and u > xm ) then uw conflicts with a
frond of Lm and Rm so that uw cannot be embedded without creating an
odd cycle in Fm . So G is non-planar in this case. If case 2 or 3 applies,
then SwitchSides is executed. uw conflicts with the fronds to `w and rw .
Suppose first that case 2 applies. Then `w belongs to a component Fi where
i < m and rw belongs to Fm . The components Fi+1 , . . ., Fm contain no
fronds in LF which conflict with uw. Therefore they all conflict with uw
due to fronds in RF . Two components Fm and Fm−1 are merged when uw
conflicts with both of them. Therefore the components of F G are always
connected. They are merged by SwitchSides into one connected component
(while um−1 > `w do) since they all conflict with uw, and then the resulting
Lm and Rm are exchanged. This is done by exchanging the portions of the
linked lists LF and RF that belong to Lm and Rm , using um , vm , xm ,
and ym as pointers into the linked lists. At this point in the algorithm
m = i + 1. It may be that Fi is a component which conflicts with uw in
both Li and Ri . The algorithm is now in a position to determine this by
adjusting the values of `w and rw , and comparing them with w. The new
value of `w is the previous value of rw , since the frond associated with rw
has been moved to the left. The new value of rw is the smaller endpoint
of the frond of RF immediately preceding Fm . Finally Fm and Fm−1 are
merged.
If case 3 applies the proof is very similar, except that right and left
are interchanged. Thus we can conclude that if G is planar, eventually all
components Fi which conflict with uw in one of Li or Ri will be aligned so
that uw can be embedded. The theorem follows.
Notice that the proof that the algorithm works when |L(T )| = 1 does
not require rule 4.2. When |L(T )| = 1, the ordering of the adjacency lists
is sufficient if just the L1 -value is used to order the fronds.
6. Graph-Theoretic Analysis of the Algorithm, |L(T )| > 1.
21
The algorithm EmbedBranch in section 3 uses a procedure EmbedFrond
when a frond uv is encountered while visiting vertex u. The action of
EmbedFrond(u, v) is described in section 5. Now when a branch Bu (v) is
encountered the entire branch must be embedded either on the left or right
of T . The branch behaves very much like a frond uw where w = L1 (v).
If it is not possible to embed a frond uw then it will not be possible to
embed Bu (v) either. This can be tested by EmbedFrond(u, w). If uw can
be embedded, say on LF , then the fronds xy of Bu(v) for which y < u
must also be embedded in LF . Those for which x, y ≥ u can be embedded
either in LF or RF . There is a very simple device which can be used to
accomplish this. Suppose that Bu (v) is to be embedded in LF . Embed a
frond uw in LF , where w = L1 (v). Mark uw as a “false frond”. Embed
a “branch marker” uu in RF . Mark it also as “false”. Place uw and uu
in the same component Fm of F G. The purpose of the false frond uw
is to determine whether there is room to embed Bu (v). The purpose of
the branch marker uu is to force all fronds xy of Bu (v) for which y < u
to the same side of T . The algorithm EmbedBranch can be modified by
replacing the group of statements beginning “if b(v) = u then begin” with
the following.
if b(v) = u then begin
{ uv begins a new branch at u }
w := L1 (v)
EmbedFrond(u, w)
if EmbedFrond is unsuccessful then Exit
mark uw as a false frond
embed a branch marker uu on the side opposite to uw,
in the same Fm
end
The definition 5.1 of conflicting fronds is not sufficient when |L(T )| > 1.
We refer to fronds, false fronds, and branch markers collectively as fronds.
6.1 Definition. Let Bu (v) be a branch with branch marker uu and false
frond ut, where t = L1 (v). Let xy and wz be any fronds where x > y,
w > z, and x ≥ w. Then:
1) The branch marker uu conflicts with the corresponding false frond ut;
2) xy and wz conflict if b(x) = b(w) and x > w > y > z;
3) xy and uu conflict if b(x) = u, y < u, and xy ∈ Bu(v).
The frond graph is denoted F G. Its vertices are the fronds of G. Two
fronds are adjacent if they conflict.
If Bu(v) is as above, with false frond ut embedded in LF , then any
frond xy of Bu (v) for which y < u will conflict with the branch marker uu
22
in RF . It will therefore be embedded in LF , as part of Fm . A frond xy
for which x, y ≥ u will not conflict with uu. It will be embedded in either
LF or RF . No frond xy of Bu (v) will conflict with the false frond ut since
b(x) 6= b(u). If it becomes necessary to exchange the bipartition of Fm , the
entire branch will automatically be switched.
The easiest way to mark a frond false is to make one of its endpoints
negative, use (−u, t) and (−u, u) instead of ut and uu. The program knows
that a negative value indicates a false frond. They become sentinels in the
linked lists LF and RF . In a similar fashion Hopcroft and Tarjan use “end
of stack markers” to delimit the branches of T . The false fronds are handy
because they contribute to the proof that the algorithm works. They form
part of the frond graph, which is still bipartite, and can be used to find odd
cycles when G is non-planar.
6.2 Lemma. If G is planar, then F G is bipartite.
Proof . If |L(T )| = 1, definition 6.1 reduces to 5.1 and the result follows by
5.2. We claim that (LF, RF ) is a bipartition of F G. If it is not, then LF ,
say, will contain conflicting fronds uv and xy, where u ≥ v and x ≥ y.
Case 1. uv and xy are true fronds.
Since b(u) = b(x), uv and xy are in the same branch, say Bz (w), where
w = b(u). The first frond st of the branch together with the path in T
connecting s to t creates a cycle C such that uv and xy are crossing fronds
both inside C. But G is planar, a contradiction.
Case 2. uv is a false frond, xy is a true frond.
As in case 1, uv and xy are in a branch Bz (w), where b(u) = b(x) = w,
and a cycle C is given. u is the branch point of a branch Bu (q) which
contains a frond to v = L1 (q). This creates two crossing paths inside C, a
contradiction.
Case 3. uv is a branch marker, xy is a true frond.
Since b(x) = u, xy is contained in a branch B at u. The branch marker
uu corresponding to B must be in RF , since xy ∈ LF . So uv must be the
branch marker for a different branch, a contradiction.
Case 4. uv and xy are false fronds.
This case is almost identical to case 2.
Case 5. uv is a branch marker, xy is a false frond.
This case is almost identical to case 3.
Case 6. uv and xy are branch markers.
Since b(u) = x and b(x) = u, we must have u = x = 1, a contradiction.
Knowing that F G must be bipartite, an odd cycle in F G indicates that
G is non-planar. This is the key to proving that the algorithm works. An
odd cycle may involve false fronds or branch markers, but still proves that
G is non-planar, by lemma 6.2.
23
When the frond graph is defined in this way, it is not necessary to
introduce segments, as in [13] or [19]. Rather the embedding consists of
the tree T , together with its fronds. F G is always bipartite. Similarly the
branch points b(v) together with the paths Pv preclude the need to define
a tree of paths as in [19].
If |L(T )| > 1, pick v ∈ L(T ) where v is the last leaf of T visited by
EmbedBranch. Let u = b(v) and let r denote the root of T . Then Pv is a
vu-path, and u 6= r. Let uw be the stem of the branch at u containing v,
and set x = a(v). Let s = L1 (v). This is illustrated below.
s
u
w
z
x
v
Fig. 9, The branch Bu (w) containing v ∈ L(T ), `(Pv ) > 1.
The proof that the algorithm will embed G when |L(T )| > 1 is based on
a decomposition of G into Bu (w) and G − Bu (w), that is, T is decomposed
into Pv and T − Pv , as in lemma 3.2. Construct a new graph G0 from G
as follows. Contract all the edges between s and a(u) into a single vertex
t, and delete all vertices and edges except t and those of Bu (w). This is
illustrated in Fig. 10. If G is planar, then so is G0 . A DF-tree T 0 of G0
consists of a single path from v to t. Any frond at in G0 corresponds to at
least one frond ay in G, where s ≤ y < u.
6.3 Lemma. If G is planar, then G0 can be embedded so that all fronds
to t are on the same side of T 0 .
Proof . Let F G0 denote the frond graph for G0 , with components F1, F2 , . . .,
Fm , where m ≥ 1. If it is not possible to embed G0 with all fronds to t on
the left side of T 0 , then exchange the bipartition of each Fi as necessary, so
that all fronds yt are on the left, where y > a and a is as small as possible.
So there is a frond at ∈ Fi on the right side, for some i, which cannot be
24
moved to the left by exchanging the bipartition of Fi without bringing a
frond bt to the right, where b > a. Choose b as small as possible. Then
there are then no fronds to t between a and b. Let P be a shortest path in
Fi connecting at ∈ Ri to bt ∈ Li . Then `(P ) is odd, since at and bt are on
opposite sides of T 0 . P consists of a sequence of fronds bt, u1 v1 , u2 v2 , . . .,
uj vj , at, where j = `(P ) − 1 is even and each uk vk conflicts with uk+1 vk+1 .
This is illustrated with j = 2. Without loss of generality we can choose
u1v1 so that v1 is as small as possible. Similarly we can then choose u2 v2
so that v2 is as small as possible, and so forth, without changing `(P ).
s
t
u
v1
b
v
a
v2
u
u2
v1
u1
b
v
a
v2
u2
u1
Fig. 10, G0 and G, b → s
Case 1. bt in G0 has a corresponding frond sb in G.
Let x1 be the vertex with a(x1) = b and let z be the vertex with a(z) = a.
In G, Bb (x1) is a type II branch which precedes the frond sb in Adj[b].
Therefore N (x1 ) is even, by rule 4.2. There are no fronds to t between a
and b, so in G, v1 = L2 (x1 ). Similarly v2 = L2 (x2 ), where a(x2 ) = v1 ,
and so forth, until we reach vj with a(xj+1 ) = vj . We have L1 (xk ) = s
and L2(xk ) = vk = a(xk+1 ), for k = 1, 2, . . . , j. By definition 4.1, N (xk )
depends on N (xk+1 ).
Suppose first that a → y, where s < y < u. Then N (xj+1 ) = 1, by
4.1. It follows that N (xj ) = N (xj+1 ) + 1, and so on, until N (x1 ) = j + 1,
which is odd, a contradiction.
We conclude that s is the only vertex < u adjacent to a. So as is
a frond. The branch Ba (z) precedes as in Adj[a], so N (z) is even, by
rule 4.2. We claim that N (z) 6= 0. Now L2 (z) = vj so N (z) depends on
N (xj+1 ). The only way that N (z) = 0 would be possible is if the last
option in item 3.1 of definition 4.1 applied. But this requires that a 6→ s,
25
a contradiction. So N (z) is even and 6= 0. Vertex xj lies between uj and
a, so L2 (xj ) = L2 (z) = vj . The last option in item 3.1 of definition 4.1
applies, so that N (xj ) = N (z), which is even. N (xj−1 ) depends on N (xj ).
Either N (xj−1 ) = 1 or N (xj−1 ) = N (xj ) + 1, in both cases an odd number.
There are no fronds to t between a and b, so we eventually must have
N (x1 ) = N (z) + j − 1, an odd number, which is impossible.
Case 2. bt in G0 corresponds to by in G, where y > s, s 6→ b.
In this case, the path P in F G0 corresponds to a path of odd length in F G
connecting by to the frond aw corresponding to at. Since by and aw both
conflict with the branch marker uu, an odd cycle exists in F G. By lemma
6.2, G is non-planar. The example in Fig. 11 shows a K3,3 homeomorph in
G when j = 2. The dashed portions of T are not part of the K3,3 . The dark
and light shading of some of the nodes indicates the bipartition of K3,3.
s
u
v1
v2
u2
b
v
a
u1
Fig. 11, G, b 6→ s
6.4 Lemma. Let |L(T )| > 1. If the algorithm runs to completion then G
is planar.
Proof . The embedding is built up by adding one frond at a time to either
LF or RF . If a frond uv conflicts with fronds of both LF and RF in Fm ,
then F G contains an odd cycle, so that G is non-planar, and the algorithm
stops. If no odd cycle is created, uv will be embedded inside a face of the
existing embedding.
6.5 Theorem. Let G be a 2-connected planar graph. Then the algorithm
will embed G in the plane.
26
Proof . By induction on n and |L(T )|. By theorem 5.3, we know that the
theorem is true when |L(T )| = 1, for all n. Let n > 3 and suppose that
the theorem holds for all graphs with fewer than n vertices. Let k > 1 and
suppose that the theorem is also true for all graphs on n vertices whose
DF-tree T satisfies |L(T )| < k. Let G have n vertices with |L(T )| = k.
As in Fig. 9 above, r denotes the root of T , v ∈ L(T ) is the last leaf of
T visited by EmbedBranch, u = b(v), a(w) = u, x = a(v), and s = L1 (v).
Bu (w) is the last branch visited at vertex u. Let z be the first vertex
of Adj[u]. uz is either a frond or the stem of a branch. In either case
wtu [z] ≤ wtu [w] and there is a cycle C containing s, u, and z, such that
the entire branch Bu (w) is embedded inside C.
Case 1. `(Pv ) > 1.
Contract edge vx in G and in T to get a graph G00 with tree T 00 . If G is
planar then so is G00 . T 00 is a DF-tree of G00 , and the weights of edges in
G00 are the same as in G. Therefore the ordering of the adjacency lists is
the same in G00 as in G, except possibly at vertex x, which is now a leaf.
G00 has n − 1 vertices, so the algorithm will embed G00 . Each frond will be
placed in LF or RF . When the algorithm is given G as input, the execution
of EmbedBranch(w) called from u is exactly the same as in the graph G0
of lemma 6.1, except that fronds to vertices < u must be embedded on
the left side. By lemma 6.1 this can be accomplished by exchanging the
bipartition of several Fi , if necessary. This is what the algorithm will do.
Because of the branch markers all fronds to vertices < u will always be on
the same side of T . Therefore Bu (w) will be successfully embedded. The
remaining execution of the algorithm is exactly the same in G as in G00 .
The algorithm cannot know whether the branch Bu (w) just embedded was
contracted as in G00 or not, the fronds to vertices < u are identical in both
cases. So if `(Pv ) > 1, G will be successfully embedded in the plane.
s1
s2
w
u
v
Fig. 12, The branch Bu (v) when `(Pv ) = 1
27
Case 2. `(Pv ) = 1.
In this case a(v) = b(v) = u. Let w = b(u), s1 = L1 (v), and s2 = L2 (v). Use
induction on m = deg(v). If m = 2, then we can replace the branch Bu (v)
with a single frond us1 . The resulting graph has n − 1 vertices, so will be
successfully embedded. Given G, the algorithm will call EmbedBranch(v)
at vertex u before embedding the frond vs1 . The rest of the execution
will be identical, except that LF and RF will contain an additional false
frond us1 and branch marker uu, which will not affect the execution. So G
will be embedded when m = 2. If m > 2, construct a new graph H from
G by replacing Bu (v) with a frond us1 and a new branch B containing v
together with the remaining m − 1 fronds at v. So in H, deg(v) = m − 1,
and the frond us1 will precede the branch B in Adj[u]. By the induction
hypothesis, H will be successfully embedded. We compare the execution of
the algorithm given H and G. In H there are fronds ww, uu, us1 , us2 , and
vs2 . In G there are ww, uu, us1 , vs1 and vs2 . If uu and ww are embedded
on the same side in H, then the execution of the algorithm in G will be
the same as in H except that all the fronds at v will be embedded by a
recursive call EmbedBranch(v). So if uu and ww are on the same side in
H, G will be successfully embedded.
If s1 , s2 < w, the frond us1 and the false frond us2 in H both conflict
with ww. Since us2 conflicts with uu, we must have uu and ww on the
same side, so G will be successfully embedded.
If s1 < w but s2 ≥ w, then us1 conflicts with ww in H, but us2 does
not. If uu and ww are in different components of the frond graph for H,
then we can switch the component containing uu to put uu and ww on the
same side. If they are in the same component on opposite sides, then F H
contains a path P of odd length connecting uu to ww. This path will also
occur in G. But in G, us1 conflicts with both uu and ww, creating an odd
cycle in F G, a contradiction. It follows that the result holds when m > 1.
By induction, the theorem holds for all values of |L(T )| ≥ 1 and all n.
Once it has been determined that G is planar, a planar embedding must
be constructed. This consists of two parts, 1) constructing a combinatorial
embedding, and 2) constructing a geometric embedding of G. A geometric
embedding is the actual layout, representing the vertices as points in the
plane, and the edges as non-crossing curves in the plane. The combinatorial
embedding is given by the cyclic order of edges at each vertex [18]. Read’s
algorithm [18] is one method that can be used to construct a geometric
embedding once a combinatorial embedding is given. A method of finding a
combinatorial embedding is given by Nishizeki and Chiba [6,17], using PQtrees. The Hopcroft-Tarjan algorithm as presented here actually provides
the geometric embedding, almost for free.
28
6.6 Theorem. If G is planar, the linked lists LF and RF completely
determine a combinatorial embedding of G.
Sketch of proof. It is clear that a “walk-around” traversal of T together
with its fronds determines the cyclic order of the edges at each vertex in G.
If |L(T )| = 1, T is a single path, and the walk-around is determined by the
order of the fronds on LF and RF . If |L(T )| > 1, let v ∈ L(T ) be the last
leaf visited by the algorithm, let u = b(v), w = L1 (v), and use induction.
The false frond uw and branch marker uu will appear on opposite sides of
T . A traversal of LF and RF will encounter the branch marker uu and
corresponding false frond uw, and “walk around” the branch. It is fairly
straightforward to develop an algorithm which will traverse the lists LF
and RF and directly assign the cyclic ordering of the edges. The details
will be given elsewhere.
If G is non-planar it contains a homeomorph of a Kuratowski subgraph K3,3 or K5 . K5 occurs only in very exceptional circumstances (Kelmans[14]), so that if G is non-planar, we usually want to find a K3,3 homeomorph in G. This problem is addressed by Williamson [19,21]. It is not
too hard to see that the DF-tree T together with its branch points b(v) are
sufficient to find the paths in G which make up the K3,3 . The algorithm
for this is not presented here. It is based on finding an odd cycle in F G.
In programming the Hopcroft-Tarjan algorithm one first begins with
the LowPtDFS in order to determine L1(v) and L2 (v). The adjacency lists
are then sorted by weight. Next, the BranchPtDFS can be executed in
order to establish the first node in Adj[u]. Then EmbedBranch is executed
in order to construct the lists LF and RF which determine the combinatorial embedding of G. An alternative is to combine BranchPtDFS and
EmbedBranch into a single DFS, which is easy to do. In either case it is
obvious that the algorithm has linear running time, since a DFS considers
each edge of G exactly twice.
Acknowledgement. I would like to thank Almira Karabeg for bringing
the book by Williamson [19] to my attention.
References
1. A.O. Aho, J.E. Hopcroft, and J.D. Ullman, The Design and Analysis
of Computer Algorithms, Addison-Wesley, Reading, Massachussetts,
1974.
2. Giuseppe Di Battista, Peter Eades, Roberto Tamassia, and Ioannis G.
Tollis, “Algorithms for Drawing Graphs: an Annotated Bibliography”,
preprint, 1993.
3. J.A. Bondy and U.S.R. Murty, Graph Theory with Applications, American Elsevier Publishing, New York, 1976.
29
4. K.S. Booth and G.S. Lueker, “Testing for the consecutive ones property, interval graphs, and graph planarity using PQ-tree algorithms”,
JCSS 13 (1976), 335-379.
5. Gary Chartrand and Ortrud Oellermann, Applied and Algorithmic
Graph Theory, McGraw-Hill, Inc., New York, 1993.
6. Norishige Chiba and Takao Nishizeki, “A linear algorithm for embedding planar graphs using PQ-trees”, Journal of Computer and System
Sciences 30 (1985), 54-76.
7. G. Demoucron, Y. Malgrange, and R. Pertuiset, “Graphes planaires:
reconnaissance et construction des représentations planaires topologiques”, Rev. Francaise Recherche Opérationelle 8 (1965) 34-47.
8. Narsingh Deo, Graph Theory with Applications to Engineering and
Computer Science, Prentice Hall, Inc., New Jersey, 1974.
9. Shimon Even, Graph Algorithms , Computer Science Press, Maryland,
1979.
10. L.R. Foulds, Graph Theory Applications, Springer-Verlag, New York,
1992.
11. Alan Gibbons, Algorithmic Graph Theory,Cambridge University Press,
Cambridge, U.K., 1985.
12. Ronald Gould, Graph Theory, Benjamin/Cummings Publishing, Inc.,
Menlo Park, California, 1988
13. John Hopcroft and Robert Tarjan, “Efficient planarity testing”, JACM
21 (1974), 449-568.
14. A.K. Kelmans, “Graph planarity and related topics”, Rutcor Research
Report #28-92, Rutgers University, New Brunswick, New Jersey, 1992.
15. A. Lempel, S. Even, and I. Cederbaum, “An algorithm for planarity
testing of graphs”, in Theory of Graphs, International Symposium,
Rome, 1966, Ed. P. Rosenstiehl, Gordon and Breach, New York, 1967.
16. James A. McHugh, Algorithmic Graph Theory, Prentice Hall, Inc.,
New Jersey, 1990.
17. T. Nishizeki and N. Chiba, Planar Graphs: Theory and Algorithms,
Annals of Discrete Mathematics, vol. 32, North Holland, Amsterdam,
1988.
18. R.C. Read, “A new method for drawing a planar graph given the cyclic
order of the edges at each vertex”, Congressus Numerantium 56 (1987),
31-44.
19. S. Gill Williamson, Combinatorics for Computer Science, Computer
Science Press, Maryland, 1985.
20. S.G. Williamson, “Embedding graphs in the plane — algorithmic aspects”, Annals of Discrete Mathematics 6 (1980), 349-384.
21. S.G.Williamson,“Depth-first search and Kuratowski subgraphs”,JACM
31 (1984), 681-693.
30
View publication stats
Download